DefectForge β foreign-body detector (YOLO11s)
Trained only on synthetic data generated and labelled by the DefectForge swarm (see the dataset): SDXL images β YOLOE labels β geometric + Qwen-VLM Critic verification β YOLO11s fine-tune (60 epochs, imgsz 640). Zero human labels in the training set.
Measured results (read the caveats)
| Eval set | mAP50 | mAP50-95 | Precision | Recall |
|---|---|---|---|---|
| Held-out synthetic val (7 images) | 0.7188 | 0.5907 | 0.7497 | 0.5716 |
| Seeded/composited test (24 images, 31 instances) | 0.3731 | 0.2855 | 0.4246 | 0.451 |
Caveats. The val split is synthetic data from the same generator as the training set. The seeded test composites defect cutouts onto fresh synthetic backgrounds β a proxy for new-scene generalisation, not field data and not a real-world performance estimate. Validate on real line imagery before any production use.
Provenance & licences
DefectForge is a port/integration of aionVIS (MIT) by 0xAthkot, run on molab (NVIDIA RTX PRO 6000) with the swarm's brain served by a self-hosted Qwen3.8-27B endpoint. Ultralytics YOLO is AGPL-3.0; SDXL and YOLOE carry their own licences.